How does Deepseek make money? Whats their business model

[-]

KeyPossibility2339@reddit

They charge for API. Chat is free but often rate limited, normal chat chat doesn’t cost a lot < 5k tokens

[-]

ragegravy@reddit

normal chat chat doesn’t cost a lot < 5k tokens

how can you check how many get used?

[-]

Roughly speaking 1 token is 3/4 word. Meaning you will cross 5000 tokens combined input + output if the combined word count exceed 3500 words. Only time I cross 5k tokens is when I paste my codebase in the chat.

[-]

huzbum@reddit

Yeah, but keep in mind, as the conversation continues, the context length builds up like a snowball.

Unless there is caching, with each message, the entire conversation is re-processed as input tokens with each message. So if you have 5000 tokens between you and the llm, and you say "thanks" then it will process all 5000 tokens from the entire conversation just to say "you're welcome" (plus however many it wastes thinking about how to reply.)

[-]

KeyPossibility2339@reddit

Yes, you’re right. Thanks for this reply, that reminded me to cache tokens in my application.

[-]

BoJackHorseMan53@reddit

Why does America charge for chat?

[-]

Acrobatic-Original92@reddit

all the upset american incels downvoted ur shit lmao

[-]

Direspark@reddit

Don't most American AI companies have free tiers...?

[-]

BoJackHorseMan53@reddit

Chinese AI chat are completely free with no way to pay.

[-]

entsnack@reddit

Sure? https://www.alibabacloud.com/help/en/model-studio/

[-]

BoJackHorseMan53@reddit

Are you retarded? America has a $20 subscription with limits AND paid API. Chinese companies on the other hand provide unlimited chat usage with no limits, but they charge for api.

[-]

WutsAWriter@reddit

The other 52 Bojack Horsemen are probably very disappointed in your bad attitude.

[-]

entsnack@reddit

Probably just annoyed at bro saying "America" like an average collectivist wumao. Could've just said OpenAI. Not everyone thinks in terms of nations and a proud-a-f citizenry (ask Americans what they think of America, you'll hear "shithole" frequently).

[-]

BoJackHorseMan53@reddit

I feel the same kinda weird when I see people say China when referring to Alibaba, so I started doing the same.

[-]

WutsAWriter@reddit

I’m American, and I have very specific opinions on this place, yes lol.

[-]

Dimi1706@reddit

Shorting Nvidia before every release

[-]

WordTrap@reddit

Does that work out? I just host my Nvidia and am up 52.9%. I believe american AI companies will buy more GPU’s to compete with

[-]

WordTrap@reddit

All big tech companies are bulding massive data centers and building powerplants. Billion of GPU’s are being bought and NVIDIA currently holds the AI market because of CUDA. However, GOOGLE is going the win the AI race because they do not depend on NVIDIA and everybody else is.

[-]

tgreenhaw@reddit

Amazon, Microsoft, and Tesla/X make their own chips for AI. AMD now has contenders for inference.

Nvidia has first mover advantage though.

[-]

cuolong@reddit

If Google is going to win, it will be because of TPUs but also because of their massive amount of data they can access, their large amounts of talent, large amount of capital and the ready-made structure they can inject agentic AI in the Google Suite to ensure widespread adoption. No other company has a confluence of all four except for maybe OpenAI combined with Microsoft.

[-]

UsernameAvaylable@reddit

This no joke, they likely made a few dozen billions there seeing that they are a quant and basically WERE insiders on this.

[-]

ohHesRightAgain@reddit

In a sane world, Nvidia would benefit from every single open LLM release... benefit more than anyone.

[-]

Django_McFly@reddit

Their stock is at ATH. One any given day, they're the most valuable corporation in the world. The world is sane in that regard.

[-]

SelectionCalm70@reddit

this +1

[-]

InspectorOk6205@reddit

They donot have any memory system , and hence if you want to build something useful , mostly API is required. Though they are providing it at very economical costs too compared to others and have quality models open sourced . Hence I think they are much more reliable.

[-]

handsoapdispenser@reddit

You can ask that of every LLM company. Deepseek is at least known for producing a model at greatly reduced cost. Major US companies spent just massive amounts of money on training and running their inference workloads so that despite substantial revenue they are deep in the red.

[-]

dexterlemmer@reddit

Deepseek trains near frontier foundation models. Not frontier base models. There's a many orders of magnitude cost difference. Someone like Alibaba may one day be able to compete with the US companies but currently, nobody outside the US does. Of course, if the Chinese can stay ahead in innovation sufficiently to make up for the knowledge gap of their models, they may not need to. But if they manage to sufficiently destroy the business models of the US giants, that in itself will possibly grind progress to a halt. Though, I don't think they would quite manage it. At worst, they'll force Tesla to buy xAI or provide a way for xAI to remain profitable but destroy OAI, Alphabet and Anthropic. Making the World stuck with only xAI at the frontier. But that probably also won't happen. Enterprise inference is getting very profitable very fast.

[-]

No_Efficiency_1144@reddit

They are frontier lab so their business model is discover AGI and then become like Mr Monopoly

[-]

-dysangel-@reddit

their founder is a billionaire I think, so he's already Mr Monopoly.

[-]

marcoc2@reddit

Seems you guys don't know how things work on China. Yes, there are billionaires, but they have many devices to avoid contradictions of capitalism.

[-]

cristoper@reddit

If they have billionaires, I don't think their measures to avoid the contradictions of capitalism are working very well...

[-]

marcoc2@reddit

And still, people call them communists and authoritarian. There are far fewer problems with having billionaires who pay their taxes and comply with state laws and plans

[-]

Relevant-Yak-9657@reddit

To be fair, one good doesn't negate their bads.

[-]

Minimum_Humor5417@reddit

To be fair

Your "fairness" must come from the same factory exporting "freedom".

[-]

Relevant-Yak-9657@reddit

Replace it with “on the other side”. It doesn’t really matter to me, my point stands regardless.

[-]

MoffKalast@reddit

China is doing parkour, seamlessly combining the worst parts of both capitalism and socialism.

[-]

Saerain@reddit

Jesus Christ.

[-]

Gimme_Doi@reddit

i think he meant monopoly as in monoploy over and above google/OAI etc

[-]

Material_Policy6327@reddit

Double monopoly!

[-]

-dysangel-@reddit

Duopoly?

[-]

No_Efficiency_1144@reddit

Yeah the market dynamics

[-]

Mindless_Parking5473@reddit

AGI is just marketing to get funding, the product is the thing they are currently making.

[-]

No_Efficiency_1144@reddit

Yes pretty much

[-]

acadia11@reddit

You think GAI is only coming from one group. Doubtful , there will be many players just like there are now. And then some form of consolidation … IMHO as communication between machines becomes more standardized. Just my take … but the mirror isn’t speaking to me today. Going to pull out the dice.

[-]

verylittlegravitaas@reddit

You clearly have no idea how any of this works.

[-]

No_Efficiency_1144@reddit

Oh I don’t actually think their business model is good but I think Mr Monopoly is the sort of model they sell to investors.

[-]

Teetota@reddit

I believe they have many sponsors interested in having an alternative to western LLMs, Chinese government among them

[-]

badmoonrisingnl@reddit

From what I know, deepseek is used in the automotive industry in China. China makes 70% of the EV cars worldwide. I am not sure if deepseek is used for self driving but it's used for driver interaction.

You should see some videos on it. It's way more advanced than what we have in the west. In fact the CEO of Ford USA had mentioned it in a recent interview how incredibly advanced it is.

[-]

Southern-Chain-6485@reddit

Deepseek (the full model, not the distilled low parameter models which aren't really deepseek) requires a huge amout of computing power to run, it's not running inside a car.

Is it possible that Chinese automakers are using it as part of their R&D? Sure. But it's not running inside a car's computer.

[-]

badmoonrisingnl@reddit

I'm sure they are not running the full model. But when asked why Ford doesn't have these advanced capabilities like the Chinese do, his answer was because we didn't go that way, we (the USA) didn't go automotive.

Newer cars here do have chat GTP now but it doesn't do what Chinese cars do. Like you order something and pay for it in your car, it knows who is driving because it uses face recognition (my car uses Bluetooth but doesn't know if it's me driving or my wife).. I don't live in China so I only can tell you what I have been told.

[-]

remedy-tungson@reddit

Most of revenue of them come from contract with other Giant companies in China.

[-]

Amgadoz@reddit

They sell LLM Inference through their API. Their latency is relatively high which means their costs aren't that high.

[-]

GFrings@reddit

Impress the CCP enough to get another grant next year. I know it sounds jaded and possibly even racist, but people need to look closer at this company. They have very clear ties to the Chinese government and their founder is now leading the national ai strategic effort. Maybe against his will, I hope he's alright.

[-]

Sketch0z@reddit

Yeah, how awful, being paid by your nation's government to do something you are deeply interested in... how unbearable

/s

[-]

jgjl@reddit

So you are telling me that all the US companies who “donated” money to their governments parties and get pushed by them do not have ties to their government?

[-]

Mediocre-Method782@reddit

A daughter of one of the US Generals on the OpenAI board posted this

[-]

Sketch0z@reddit

Being Chinese... I imagine their government subsides a lot about their R&D.

Neat thing about attempting communism is that if your dreams happen to align with the government's, you don't need to think about profit. You have to think about making China the greatest at whatever the thing is. In this case, tech and AI dominance.

[-]

megadonkeyx@reddit

they are backed by a hedge fund apparently, they used GPUs for that and kindof drifted into AI.

[-]

MaxKruse96@reddit

the only correct answer here. just how qwen is good because its alibaba backed, deepseek is good because of the hedge fund

[-]

Minute_Attempt3063@reddit

From what I remember, they are a crypto hedge fund. The CEO or one of the people within the company, had a interest in ai, and then they made what we have now.

Even with the API costs, i think they loose a lot of money, but maybe that doesn't matter for them... As it is a side project for them

[-]

121507090301@reddit

Even with the API costs, i think they loose a lot of money

Didn't they say they actually made money?

[-]

Peach-555@reddit

They make money on the inference they sell yes, they made a post earlier this year about how they had a theoretical 8x profit margin. I don't think they published the actual numbers, but they put out an article that described how their machines that could be rented for $x per hour could output tokens sold for $9x per hour.

If I remember correctly something like 100 million tokens for \~$10 which they could sell for \~$100.

[-]

Electronic_Sign_322@reddit

im surprised they can sell inference at a profit… maybe cheap electricity costs and hypothetical mass produced lower tier chinese gpu’s (idk)

[-]

Minute_Attempt3063@reddit

If they did say that, then I stand corrected.

[-]

nullmove@reddit

Yeah they were already doing a lot of ML for algorithmic trading.

Also there are some rumours going around that they will pursue IPO in fall (these don't seem reliable nor do I believe them myself).

[-]

Electronic_Sign_322@reddit

ipo would surprise me. i figured they’d stay employee owned. I guess we’ll see

[-]

SAPPHIR3ROS3@reddit

All of deepseek started as a S I D E P R O J E C T

[-]

aero-spike@reddit

API

[-]

El_Guapo00@reddit

Your data ....

[-]

entropyideas@reddit

They should paying you to train AI, LOL. Eventually all the trained data going be sold for billions.

[-]

yani205@reddit

Deepseek to the hedge fund owner is like WhatsApp to Zuckerberg - it was unprofitable, but that will change

[-]

Amazing_Athlete_2265@reddit

Your data is worth $$$

[-]

No_Conversation9561@reddit

not my data though.. it’s worth shit

[-]

No_Efficiency_1144@reddit

Qwen giving 1,000 free daily uses of Qwen Coder is prime example.

Still worth using BTW for non-sensitive data. This trade-off is something that can benefit both parties. The alternative is Claude pricing.

[-]

Recoil42@reddit

Qwen giving daily use of Qwen Coder is a marketing expense. They want enterprise customers.

[-]

No_Efficiency_1144@reddit

Yes you’ve got the spirit.

I use local loads but in the range of 0.006B to 0.6B or so

[-]

Recoil42@reddit

I'm telling you it really isn't a data thing. User data from Qwen Coder is more or less useless, especially for an organization as big as Alibaba (Qwen). What they're after is enterprise business. Big fish who'll spend millions on Qwen Coder for all their employees.

[-]

No_Efficiency_1144@reddit

We are hitting a “data wall” though I think the cost of data will go up.

When I follow open source coding LLM training projects they are always complaining about not enough data.

[-]

Recoil42@reddit

There is no data wall. No such thing exists.

The wall in ML/AI right now is all architectural, and with coding and LLMs in specific, we already know we can use synthetic data to improve performance.

[-]

No_Efficiency_1144@reddit

Sure but synthetic is drastically more expensive than when they just took big pre-made datasets off the internet in the GPT 3 days

[-]

Recoil42@reddit

Not at all — synthetic data is not only dirt cheap, but basically the happiest path there is to RL. See R1-Zero.

[-]

No_Efficiency_1144@reddit

For text-based LLMs synthetic data pipelines need a lot of RL runs to get right, a single RL run for a 500B LLM is currently around $500,000. For things like robotics and driving, the synthetic data pipelines are pretty much the same cost as developing a full video game or CGI movie. I’ve been looking at the synthetic data startup opportunities and sadly you can’t really get it done for even $50,000 seed funding. It is more like $1,000,000 to $10,000,000 mark at the moment.

[-]

ivxk@reddit

In the grand scheme of things the marketing is worth so much more than the data.

Most users won't ever hit the daily cap, converting a single one to a paying customer is worth tens of free users.

And it's not like those companies aren't beyond dumping millions just to capture mindshare and users to then figure out how to get money out of them way later on. We're talking about Alibaba, they're not in any way desperate for profit from qwen.

[-]

No_Efficiency_1144@reddit

Yes and also to collect training data

[-]

BoJackHorseMan53@reddit

Same as the paid AI companies btw

[-]

No_Efficiency_1144@reddit

Depends which API they don’t collect from the enterprise tier APIs

[-]

BoJackHorseMan53@reddit

It's impossible not to collect data because that's how LLM inference works. They say they don't collect data but do you trust the companies known to violate copyright of other companies?

[-]

No_Efficiency_1144@reddit

I only trust 3 specific entities- AWS, Azure and GCP enterprise accounts. I don’t trust others (Coreweave-tier probably also fine)

[-]

BoJackHorseMan53@reddit

What about Oracle cloud?

[-]

No_Efficiency_1144@reddit

I really do only trust the big 3. It’s not just about intent it’s about having really robust security including both cybersecurity software/hardware but also staff training.

[-]

BoJackHorseMan53@reddit

You don't trust Oracle? Poor guy :/

[-]

No_Efficiency_1144@reddit

Yeah it’s not the same level as the big 3

[-]

BoJackHorseMan53@reddit

Name an AI company that isn't training on your data or doing other kinds of things to make money from it, be it free or paid.

[-]

No_Efficiency_1144@reddit

Literally any enterprise API. They don’t train on enterprise tier.

[-]

Amazing_Athlete_2265@reddit

The alternative is local!

[-]

agent00F@reddit

Funny the top comment makes no attempt to answer the question, and it's wrong in any case that the data is worth more than inference cost.

Real answer is they are subsidized by previous earnings from the owner's quant fund, because Liang had a personal interest in AI since studying it in school.

[-]

Bus9917@reddit

https://www.youtube.com/watch?v=UZEn-s9mllI

[-]

nguyenm@reddit

On the surface, they sell access to their API, and the price per million input or output token is publicly available. Deepseek is a little bit more nuanced where price can vary by time-of-day demand as well as "cache misses".

What's unresolved is who's paying for the API access in enough volume to justify the cost of operations, R&D, and "growth".

[-]

vorwrath@reddit

Mostly me I think, I spent $10 on their API a few months ago. Asked R1 about 600 questions, still have $4 left!

I assume they're losing a ton of money like every LLM company and are burning through either government or private funding and resources. They'll be slurping all the submitted data for training and other purposes.

There's a strong argument to be made that having one of the best models will be very valuable in future. Once the "switch" part of the bait and switch starts, and everyone is used to leaning on AI models, what was once cheap or free will probably cost a chunky subscription. Plus there'll be enshittification and companies paying for recommendations when users ask the LLM "What is the best X" or similar queries. And shaping the answers to historical or political questions is potentially a hugely powerful and valuable tool (not naming any particular country because they will all be at it).

[-]

Southern-Chain-6485@reddit

And will the market validate the price of a chunky subscription?

[-]

robert-at-pretension@reddit

Selling Nvidia shorts 😂

[-]

Crierlon@reddit

The AGI research feeds into their hedge fund.

[-]

HarambeTenSei@reddit

They're a hedge fund that makes money from trading and AI is just a side hobby

[-]

RazzmatazzReal4129@reddit

Conspiracy theory, they use all the data they get from their API to learn insider information for their trading algorithm to use.

[-]

Django_McFly@reddit

So basically the tech version of Red Bull?

[-]

Special-Economist-64@reddit

This is the right answer . I’m surprised so many people don’t know yet.

[-]

Accomplished-Bill-45@reddit

It’s a research lab fully backed and owned by a trading hedge fund

[-]

specimen174@reddit

None of the AI platforms are 'making money' they are burning through $billions in loans to 'win the race' (which is aparently AGI or something) whoever wins gets to stay in business all the others will vanish in a puff of un-repayable debt

[-]

LostMitosis@reddit

DeepSeek is a side hustle.

[-]

czxck001@reddit

the company is a hedge fund so they can afford running the ai business at a loss

[-]

Consistent_Toe_7633@reddit

Deepseek is backed by a company called "High-Flyer Quant" (https://en.wikipedia.org/wiki/High-Flyer), which specializes in AI-driven automated stock trading. This provides Deepseek's research team with substantial resources.

On March 1, 2025, Deepseek officially published an article on Zhihu. The article’s conclusion mentioned profitability metrics: API services generated $562,027 in daily revenue with a 545% profit margin.

The claim about government cash handouts is absurd.

[-]

cheechw@reddit

People are going to ask "who uses DeepSeek API instead of OpenAI", forgetting that the entire Chinese market exists that can't access Gemini and openAI.

[-]

Consistent_Toe_7633@reddit

Right on point. Yet this very inaccessibility acts like a moat, shielding the growth of Deepseek and Qwen, both of which have moved beyond their local markets to compete alongside OpenAI and Anthropic, leading the global open-source community.

[-]

Dodokii@reddit

Governments tend to back Companies for their own agenda. Having profitable sources does not exclude government funds. That's everywhere!

That's not saying Deepseek is government funded, though, as most of the time we can not know.

[-]

fallingdowndizzyvr@reddit

Exactly. This information is widely known yet people have to make up conspiracy theories to fit their agendas.

[-]

JabootieeIsGroovy@reddit

Deepseek is a quantitative hedge fund that originally used machine learning models and other algorithmic methods to trade assets and crypto. They have an extensive AI research team and a large supply of GPUs that allowed them to create their own LLM. As an AI company, they make money mainly through their developer api. As a hedge fund, well they just have a money printer.

[-]

TwistedBrother@reddit

I periodically teach Chinese students. They all prefer DeepSeek to ChatGPT, suggesting it gets them better. That’s a lot of clientele.

Please tell me your next thought isn’t something about the party and surveillance. You should realise that Open AI is under court order to retain everything. The reality is that in China is extremely popular as a service.

[-]

2CatsOnMyKeyboard@reddit

How does any AI developing company make money? None of the American companies are profitable with regard to selling inference of the models they created as far as I know.

[-]

TangoRango808@reddit

State sponsored AU doesn’t want money it wants data and users to extract what they need IP, secrets etc

[-]

GrungeWerX@reddit

China’s playing the long game not the short game.

[-]

evilbarron2@reddit

I dunno. If Deepseek’s already integrated as a platform in China’s automotive industry then they’re already ahead of the west. Are any western frontier models being used in a comparable way in any western industry doing actual work?

Near as I can tell, OpenAI, Anthropic, Gemini and Mistral are mostly used for dicking around and porn chatbots. We’ve invested trillions already and don’t have much more than promises to show for it.

Deepseek’s ROI seems much higher than anything you can point to in the west

[-]

GrungeWerX@reddit

Um, no. You have zero idea of OpenAI’s integration into the US economy if you think that. Everybody uses chat gpt. Can’t believe you even made such a ridiculous claim.

[-]

evilbarron2@reddit

Ok, show me a company whose product depends on ChatGPT. Not one of these lane-ass apps or coding buddies, but a real-world business building real-world things and making real-world money. DeepSeek powers the non driving parts of China’s EVs: user interface, maintenance records, cabin controls, etc. what does ChatGPT do except burn tokens and build data centers?

[-]

Interesting8547@reddit

It's not free, they use our conversations to train models. Of course you can toggle that off, but most people probably don't (or even don't know). Though I don't mind that some future model might learn something from me.

[-]

Feeling-Wonder-9929@reddit

still cant see the correct answer, in fact the company behind deepseek is actually a quant cooperation, LLM was not their primary focus

[-]

tillybowman@reddit

what's a quant corp?

[-]

Feeling-Wonder-9929@reddit

Someone gave some detail, its like a jane street ig, i am not very knowledgeable for this as far as i know they are not do that for profit

[-]

tillybowman@reddit

i don't even get why i'm downvoted. anyway, thanks!

[-]

Orolol@reddit

https://www.avatrade.com/education/online-trading-strategies/quantitative-trading

[-]

OmarBessa@reddit

Huàn Fāng (aka High-Flyer) was already a successful hedge fund

if we take Deepseek for its merits, it's very likely they are using it for market analysis (i.e. getting actionable insights from every newspaper in the world) and algo trading

given the discounts on their inference, it's not like they care too much about that part of the business even though it is very profitable

[-]

SteveRD1@reddit

For every 1000 people using deepseek to ask questions of no interest to anyone else...there is probably someone having a conversation with deepseek about their employers activity they likely shouldn't be!

Got to be able to get some useful intel out of that with the right kind of monitoring in place.

[-]

BoJackHorseMan53@reddit

The trade on the stock market to make money, AI is their side business.

Just like Google sells ad space to make money and AI is their side business.

[-]

axiomaticdistortion@reddit

What they eat, I don’t know. But they drink westeners tears.

[-]

-p-e-w-@reddit

Their business model is the same as that of all other AI startups: Impressing venture capitalists so that they dump freight trains filled with money at their doorstep, on the vague promise that they will figure out how to monetize their product at some later point in time.

[-]

OmarBessa@reddit

> Impressing venture capitalists so that they dump freight trains filled with money at their doorstep

they are already bankrolled by Huàn Fāng (High-Flyer)

it's obvious from their activity that their LLMs are great for algo trading and market analysis

[-]

ColorlessCrowfeet@reddit

They aren't looking for VC money.

[-]

-p-e-w-@reddit

They’re not looking because they already have it. Their parent company is literally a hedge fund.

[-]

Jealous-Ad-202@reddit

So you just contradicted yourself. Amazing debate skills.

[-]

-LaughingMan-0D@reddit

Big ass bubble

[-]

slolobdill44@reddit

Kinda all tech startups tbh

[-]

FinBenton@reddit

Its the same as any other AI company, they are almost all just losing a lot of money and running on funds from investors in a hope for turning it to profit in the future.

[-]

Illustrious-Dot-6888@reddit

How does Meta,Google etc make money?

[-]

gyzerok@reddit

They sell ads

[-]

Illustrious-Dot-6888@reddit

Ads and YOUR data,just like Deepseek does

[-]

_thr0wkawaii14159265@reddit

You don't know that.

Anonymous data? Of course.

User identifiable data? Baseless accusation.

You just swung to the opposite of naivity, which is cynism.

[-]

FullOf_Bad_Ideas@reddit

Make god, give away god.

[-]

CCP_Annihilator@reddit

Highly profitable API

[-]

meshreplacer@reddit

I think Deepseek is a hobby for the founder. ie why not? better than throwing it away on 50 million dollar weddings to plastic looking freaks.

[-]

KernQ@reddit

I imagine there's market for producing bespoke models for industry / government.

[-]

Orugan972@reddit

I think that it's a purpose-driven activity to go beyond frontier

[-]

Skeptical0ptimist@reddit

The goal could also be to deny western companies from making money and therefore to make them fold, leaving the domain open for Chinese domination afterwards.

[-]

naiveoutlier@reddit

None of the LLM servers make money.

[-]

PutMyDickOnYourHead@reddit

Deepseek is just a fun side project for High-Flyer (quant investment firm).

[-]

Trilogix@reddit

Deep state enters chat...

[-]

Valhall22@reddit

If it's free, you're the product 🤫

[-]

IGiveAdviceToo@reddit

DeepSeek is more like a side project of a quant fund company. It doesn’t really need to make money at all ~ plus mostly funded by their billionaire owner. But for profit they earn by usage of their api to their model. For cost to performance, their R1 and V3 model really still good.

[-]

ELPascalito@reddit

I mean the LLM is paid, you pay per token, what else are they supposed to do?